AI Needs Inference, Incentives, and Institutions Around the Model

Michael Jordan does not describe himself as an AI researcher. He was trained as a statistician and cognitive scientist, and he places his own work in the tradition of machine learning systems built for real-world settings: supply chains, commerce, transportation, healthcare, finance, and large-scale economic systems.

That distinction matters because, in Jordan’s account, the current vocabulary has distorted both research and business thinking. “AGI,” he says, is “a PR term.” The older AI project, associated with symbolic reasoning and logical inference, did not produce the industrial systems that came to matter. Those came from methods developed across statistics, operations research, and related fields: decision trees, nearest neighbors, logistic regression, hidden Markov models, random forests, gradient-based methods. Long before the current LLM cycle, machine learning was already embedded in industry.

The term AI returned with force, Jordan argues, because the data became language data. A statistical box that once made predictions about supply chains, prices, demand, and logistics began producing fluent human text. If the AI problem is defined narrowly as something like the Turing test, he concedes, one might say that large language models made progress on it. But he thinks the cultural and commercial interpretation went far beyond that. The field did not merely say: here is a powerful new predictive system trained on language. It revived a story about intelligence, understanding, assistants, superintelligence, and AGI.

For Jordan, that story is not harmless. He says it has had a “distortionary effect” on research priorities, on business models, and on how younger builders understand their own options. He objects especially to the way public discussion has collapsed into two postures: exuberance about superintelligence or alarmism about extinction. Twenty- and twenty-five-year-olds, he says, are watching “thought leaders” and being offered those as the two legitimate emotional stances toward technology.

The alternative Jordan wants foregrounded is not anti-technology. He repeatedly emphasizes that the systems are useful, that builders have produced impressive artifacts, and that large-scale gradient descent has worked “more than we would have ever imagined.” His criticism is that the dominant framing is detached from the social and economic systems in which the technology acts. Machine learning systems are trained on inputs from billions of people and are meant to serve billions of people. That already makes them collective systems. To treat them primarily as disembodied minds, or as proto-agents on a path to AGI, is to miss the network of producers, consumers, incentives, data flows, privacy losses, contracts, markets, and institutions in which their real effects will be determined.

That is the premise of Jordan’s paper “A Collectivist, Economic Perspective on AI,” shown on screen during the discussion. The paper’s abstract begins from the claim that omnipresent data collection and machine learning are affecting the human world at unprecedented scale; Jordan’s spoken version is more direct. “Collectivist,” for him, means the technology is already collective in its inputs and its intended beneficiaries. “Economic” means the analysis must include money, incentives, strategic behavior, markets, and social welfare, not only architectures and benchmarks.^†

The “secretary sitting on your shoulder” image of LLM deployment is, for Michael Jordan, a narrow and unattractive model of what the technology could become. He does not deny that people may want search, summaries, or assistance at the end of a day. But he calls the always-present assistant “a dumb business model,” partly because many people will not want a constant entity whispering to them as they work and think.

His larger objection is that the assistant model is too narrow relative to the systems already being transformed by data. Healthcare, transportation, finance, commerce, and supply chains are not just settings where a user queries a model. They are multi-agent systems with many producers and consumers, asymmetric information, incentives, cooperation, competition, privacy concerns, and values being created or destroyed. Jordan wants AI work to begin from questions such as: What ecosystem does this system belong to? Who interacts with whom? At what rate? With what quality? What value is being created? Where does money flow? What jobs or creative opportunities appear?

That is why he calls his perspective “collectivist” and “economic.” “Collectivist” does not mean state planning in his usage. It means the technology is already based on collective input and intended for collective use. “Economic” means one cannot analyze the system only as a computational artifact; one must analyze incentives, markets, games, contracts, and social welfare.

He contrasts this with a Silicon Valley style of explanation that he finds underdeveloped: humans are intelligent; the brain is a computer; if one mimics and scales aspects of it, useful intelligence will emerge. Jordan thinks that story often stops before it states a goal in society. “It’ll just solve problems for us,” as he characterizes the view. He left Silicon Valley in part, he says, because he became tired of that style of talk: too much aspiration, not enough long-term intellectual structure.

Jordan also resists treating “AI safety” as a self-sufficient label. He calls it a buzzword when it is used as something placed around an opaque model without specifying the surrounding system. His alternative emphasis is predictability, input-output behavior, constraints, incentives, and interaction. The question is not only whether one can find a human-interpretable circuit inside a neural network; it is whether the larger arrangement lets people plan around the system, use it, challenge it, and avoid being harmed by it.

He uses the example of a loan denial. If a bank denies someone a loan using a large AI system trained on past data, the explanation the person needs is not a mechanistic circuit inside the model. A useful explanation might show a set of similar people, according to the embedding or representation the system uses, where some got the loan and some did not. The borrower could then see actionable differences and potentially change something. For Jordan, explanation is valuable when it supports action and interaction, not when it merely satisfies an abstract demand to peer inside the model.

He is therefore not opposed to building systems one does not fully understand. Human choices are not fully explicable either; he notes that no one can completely explain why another person picked one Airbnb over another. What matters for interaction is that people and systems become somewhat predictable, that their behavior can be modeled at the level needed to plan, coordinate, and avoid harm. Chemical engineering, in his telling, did not always understand every phenomenon in detail before useful systems were built. But it developed concepts, constraints, and system-level principles. Jordan’s complaint is that current AI discourse often has builders and intuitions, but not enough of the analogous intellectual machinery.

That is also why he resists the idea that scaling multi-agent LLM systems automatically supplies the economic layer. Scarfe frames the challenge in terms of Silicon Valley arguments that human value functions and multi-agent systems might produce the economics “for free.” Jordan’s answer is that this is not a good way to think about engineering. In chemical engineering, merely throwing materials together can produce explosions, non-viable processes, and harm. In AI, he says, similarly disruptive systems are being built with metaphors rather than social science or mathematics adequate to the setting.

Michael Jordan treats AlphaFold as a good example of both the power and the limits of modern machine learning. He admires it and distinguishes it from LLMs: AlphaFold is targeted at a particular class of problems and performs very well. But he argues that using a foundation model in science requires more than high predictive accuracy. Scientists ask new questions, often at the edge of what has previously been observed. That is where a model can be accurate overall while biased for the specific question being asked.

The example concerns AlphaFold’s roughly 200 million predicted protein structures and a biological hypothesis. In speech, Jordan describes the question as whether “quantum fluctuations” in a protein were associated with phosphorylation — whether the protein was active in the cell. The visual material on screen, drawn from the prediction-powered inference work, frames the displayed test more specifically as confidence intervals for the odds ratio of a phosphorylation event given the presence of an intrinsically disordered region, or IDR, in a protein. Jordan’s spoken shorthand was “quantum fluctuations”; the displayed figure formalized the test in terms of intrinsic disorder and phosphorylation.

The statistical structure is a two-by-two table: yes or no on phosphorylation, yes or no on the relevant disorder or fluctuation. Using only experimentally known protein structures, Jordan says, there is not enough data to reject the null hypothesis with high power. Using AlphaFold predictions at massive scale, the test has high power and rejects the null.

The problem is that the confidence interval obtained from those predictions can be extremely narrow and far from the gold-standard value. Jordan says his group found this pattern “in domain after domain.” The likely issue is not that AlphaFold is globally poor. It is that the training set may contain relatively few examples for the particular phenomenon being queried — in his telling, proteins with the relevant kind of fluctuation, which are hard to crystallize and historically less studied. AlphaFold does not, by itself, give error bars for the scientific question the researcher later asks.

The remedy Jordan describes is prediction-powered inference: use machine-learning predictions, but supplement them with a smaller amount of ground-truth data to produce valid statistical inference. The goal is to keep much of the power gained from a huge prediction set while shifting the interval so that it covers the truth in the way classical statistical inference is supposed to.

The “Prediction-Powered Inference” paper, shown on screen, defines the method as “a framework for performing valid statistical inference when an experimental dataset is supplemented with predictions from a machine-learning system.” The displayed abstract says the approach can compute provably valid confidence intervals for quantities including means, quantiles, and linear and logistic regression coefficients, and that more accurate predictions translate to smaller confidence intervals.^†

Jordan’s conceptual point is broader than AlphaFold. He distinguishes prediction from inference. A foundation model can be powerful as a predictive artifact, but scientific use requires asking: What is the estimand? What is the relevant bias for this specific question? What small amount of labeled or ground-truth data would let us correct the prediction set? Where are the error bars? Without that layer, scientists can be led astray by confident answers in precisely the places where science is most interested: the boundary of current knowledge.

He also resists saying AlphaFold “understands.” Tim Scarfe reports that John Jumper was allergic to the term and described AlphaFold as a machine that lets researchers predict and perhaps control, while humans still derive the understanding. Jordan’s response is sharper: “Why should AlphaFold understand?” He thinks anthropomorphizing intelligence and understanding is unnecessary and distracting for many problems.

His analogy comes from industrial machine learning. Around 2000, he says, Amazon was already using large amounts of data and the machine-learning methods of the day, including random forests, for supply-chain modeling. Such systems could make useful predictions about whether ships would be delayed and whether parts would arrive on time. No human could “understand” the whole supply-chain box in the ordinary sense. The question is whether the system reduces uncertainty enough to enable planning, stockpiling, logistics, and optimization. Asking whether the system understands transport and logistics is, for Jordan, mostly a media question.

The economic lens becomes concrete in Michael Jordan’s discussion of drug discovery and regulation. The standard statistical framing would ask whether a protocol controls false positives and false negatives: bad drugs should not be approved; good drugs should not be rejected. A regulatory agency, in this idealized view, is solving a statistical testing problem.

Jordan says that frame is incomplete because the data do not arrive from a neutral IID source. They arrive from self-interested pharmaceutical companies. Those companies may want to help people, but they also want money. A regulator does not directly observe all their incentives, internal evidence, or strategic choices. That turns the problem into one of incentive design.

The FDA-testing slide shown during the discussion makes the point with a stylized protocol. Bad drugs have a 5% chance of approval and 95% chance of non-approval; good drugs have 80% power and 20% chance of non-approval. The slide then contrasts two payoff regimes. If a trial costs $20 million and approval yields $200 million, the expected profit for a bad drug is negative $10 million, so the payoff regime discourages submitting bad drugs. If approval yields $2 billion, the expected profit for a bad drug is $80 million, so the regime can reward submitting bad candidates and hoping for a false positive.

The point is not that the specific slide exhausts drug regulation. It is that a nominally valid statistical test can fail as a system if the incentives induce companies to submit many bad candidates. If a drug could be used by a billion people and generate enormous returns, then even a low false-positive rate may be worth exploiting. “You just throw it at the regulatory agency,” Jordan says, and if there is a false positive, it reaches market and makes money.

The regulator’s problem is therefore not only to set a threshold but to design a system in which firms are incentivized to submit candidates they have reason to believe are good. Jordan compares this with pricing and service design in airlines: different travelers have different willingness to pay and different needs, which the airline cannot directly know. The airline offers menus of services and prices that reveal enough of those private values to make the system work. In drug discovery, the analog would be a mechanism that induces truthful, useful submissions rather than adversarial exploitation of statistical noise.

This is why Jordan says he cannot imagine AI being rolled out across society without a “deeply microeconomic perspective” accompanying machine learning. Local data will matter. Ground truth will matter. Parties will not simply give away valuable information. They will need reasons to share data, reasons to share correct data, and reasons not to behave adversarially. For AI builders and regulators, “gradient descent on data” is not a sufficient description of the system they are building into.

Michael Jordan’s three-layer data market is a deliberately minimal model, “like Bohr atom kind of things,” meant to isolate core behavior. The layers are users, platforms, and data buyers. Users provide data to platforms in exchange for services. Platforms use the data to improve their services. But many platforms also sell data, often noisily or privately transformed, to third-party buyers who want market research or behavioral insight.

The diagram shown on screen summarizes the model as a user with data $\theta \in {\mathbb{R}}^d$, multiple platforms receiving signals from that user, and a buyer interested in learning a function of the user’s data. The platforms can pass on private, noise-added data to the buyer in exchange for payment. The accompanying paper, “On Three-Layer Data Markets,” describes the same structure: users as data owners, platforms, and a data buyer, with users receiving platform services while incurring privacy loss when their data are shared with the buyer.^†

Jordan stresses that third-party data buyers are not necessarily villains. The problem is structural. Once a third layer appears, the user loses some privacy. That loss should change the equilibrium. A user may still need the platform and may not simply walk away. A platform may need revenue beyond a small transaction cut. A data buyer values less noisy data more than more private data. Those interests are aligned in some respects and conflicting in others.

The economic design question is what options should exist. A platform might offer users tunable differential privacy at some cost. One company might offer a privacy level of 0.3, another 0.7. Users who care more about privacy may choose the latter. That platform may attract more data and improve its service, but its noisier data may be less valuable to buyers, who then pay less. The system is not solved by optimizing a single objective. It is an equilibrium problem involving statistical prediction, privacy, prices, user welfare, platform revenue, and buyer value.

Jordan notes that, in the simple three-layer model, one can write equations and calculate equilibria; it is a Stackelberg game. In other cases one might simulate. But the important distinction is that many machine-learning researchers are trained to optimize and predict, whereas markets require fixed points, equilibria, Pareto frontiers, social welfare comparisons, and parameter shifts. Economists historically had equilibrium tools but not much data. Machine-learning researchers have data but often do not model equilibria. Jordan thinks the future requires these branches to meet.

His criticism of “we have all the behavioral data, so the economics is built in” is that it is too naive. He grants that data can improve on overly rational assumptions in economics and can incorporate some behavioral regularities. But data outside an economic model can still produce a mess. Social knowledge is ephemeral, contextual, and often generated in the moment. Even exabytes of historical data may miss the main details relevant to a particular decision happening now.

This is part of Jordan’s broader distinction between data and knowledge. Social knowledge includes what is available on a street in Copenhagen, at what price, for whom, under what temporary context. A system cannot simply gather enough data to know what a person will buy in the next ten seconds. Markets, cultures, and organizations let bottom-up preferences be expressed without requiring a godlike designer to encode a human value function from above. Privacy, data access, compensation, and model performance become linked once a platform functions as a market.

The same market logic shapes Michael Jordan’s comments on music and creator platforms. Spotify, in the example raised by Scarfe, can be incentivized to generate songs with AI. Jordan says he is a scientific advisor to UnitedMasters, which he describes as an alternative that lets musicians keep their work and connects them to brands and opportunities. His interest is in building a market where artists are more than sources of streamed songs that produce small payouts.

Jordan’s “Music in the Data Age” slide describes a two-sided or three-way market: dashboards can help musicians learn where their audiences are, enabling them to perform where demand exists; consumers and producers become linked; brands can partner with musicians based on affinities; the company creating the market profits by taking a cut from transactions.

Jordan’s criticism of Spotify is not that it should not exist, but that artists are paid very little and prices do not appear, to him, to be set by competitive mechanisms. He describes Spotify as close to a monopoly and says one would hope the market would respond if enough young artists conclude they are being “screwed” and move to an alternative. But he also notes that some services become monopolies quickly, raising questions for economists about regulation and market-making mechanisms.

He extends the critique to YouTube. Search, in his view, was a major technological achievement. But once Google acquired YouTube, Jordan thinks the company should have recognized that it had created more than a pointer to websites: it had created a producer-consumer market. A socially responsible design would have built a more direct economic connection between viewers and creators, so incentives could flow from audience to producer. Instead, he says, the economic relation largely went through Google, which placed advertisements beside the content, kept substantial value, and returned a modest amount to creators. Facebook, he adds, made the pattern worse.

The implication of Jordan’s argument is that creator platforms should be evaluated not only as recommendation engines, content hosts, or ad businesses. They also function as market designs. They determine who can express preferences, who captures value, whether creators can see and act on demand, whether privacy is lost, and whether incentives reward human creativity or substitute it away.

Claims about recursively self-improving, agentic AI are, to Michael Jordan, “very science fiction.” He says science fiction has social value, but he objects to its elevation into leadership rhetoric. In his view, prominent voices who warn that AI will likely wipe out humanity, or who imply that superintelligence is imminent and little remains for humans to do, are sending a demoralizing message to younger technologists.

He does not deny real risks from autonomous software. But he sees more grounded concerns in labor-capital relationships, institutional failures, platform incentives, and humans using technology badly. The central challenge, for him, is not that a recursive algorithm becomes a runaway virus. It is that society fails to design human-machine systems that improve human decision-making, reduce harmful uncertainty, and create broadly valuable opportunities.

Airplanes are his example of useful automation. Modern flight is safer, he says, in large part because of autopilots. Humans did not evolve to fly large aircraft; automation improves human capability. But the success comes from a hybrid system, not from putting “a superintelligence” in charge. Driving is a more complicated, more crowded domain, and Jordan’s conclusion is still system-level: autonomy must be considered with infrastructure, humans, other agents, and safety properties of the whole arrangement. “Just putting a superintelligence behind the wheel of a car,” he says, is “dumb.”

His positive vision is that AI should help humans with things humans are bad at, especially at the scale of billions of people. People misunderstand each other’s motivations, misread intentions, make destructive decisions under uncertainty, and build broken political systems. Jordan mentions war as a case where failures of signaling and uncertainty can lead to proactive violence. He also criticizes contemporary political decision-making as a deeply flawed human system, while noting that universities, companies, and other associations can work comparatively well.

AI, in this framing, is not about replacing humans with computers. It is about improving information flow so that humans can make the decisions they would have made if they had known enough and had better mechanisms for coordination. That is why he is “bullish” about AI in one sense and appalled by the public dialogue in another. The useful path requires mechanisms, education, goals, and institutional designs. The dominant debate, in his account, offers spectacle instead.

Jordan is careful not to dismiss the builders themselves wholesale. He says people such as Ilya Sutskever have built systems that many now use and that are changing thought. His criticism is directed at the surrounding culture: a Silicon Valley environment in which far-flung claims, neuroscience-inflected metaphors, physics language, and guru posture can attract money and status. He says figures such as Elon Musk and Sam Altman are “taking the cream off the top” of infrastructure and data built by previous generations, often without enough appreciation for the goals behind that earlier work.

What worries him is not merely hype. It is what he calls an unusual “detachment from reality”: the conversion of “25-year-old dreams” into civilization-scale programs pursued with enormous money and little clarity about the social goal. When those programs produce powerful artifacts whose purposes remain unclear, the result, in Jordan’s view, is not intellectual maturity. It is a sign that the field has mistaken the ability to build for the existence of a sufficient design philosophy.

Michael Jordan describes game theory as a mathematical discipline analogous, in spirit, to physics. If one writes down a game, as one writes down $F=ma$ in a coordinate system, the theory makes predictions. In mechanics, one integrates a differential equation and predicts a trajectory. In game theory, one writes down the strategic setting and calculates Nash equilibria, correlated equilibria, Stackelberg equilibria, sequential equilibria, or other solution concepts. One then asks whether those equilibria characterize real behavior.

That is the forward direction: given the setup, predict what happens. Mechanism design is the inverse direction. If the goal is to build a bridge, one does not merely predict a parabolic path; one works backward from the desired outcome to the design that ensures it. Mechanism design asks: what game should be designed so that a desired outcome occurs?

Contract theory is the part of mechanism design Jordan works on. It concerns interactions between asymmetric parties, where one party has private information and another wants to induce actions that depend on that information. Auction theory is another branch, where a mechanism reveals values among bidders and allocates a good accordingly. Both are ways of designing rules so that self-interested behavior produces acceptable outcomes.

A slide shown during this part of the discussion defines contract theory as a branch of incentive theory in which agents possess private information and a principal wants to incentivize actions that depend on that private information. The example on the slide is airline fares: “business fares” and “economy fares” allow a service provider to offer different prices to agents with different willingness to pay without requiring them to reveal their private values directly. The design problem is a menu of options — service and price — from which agents select.

This distinction matters for AI because a large social deployment is not merely a prediction problem. It is a design problem. If the desired outcome is fairer compensation, lower privacy loss, better drug approval behavior, reliable evidence gathering, or a functioning data market, one must design the game in which models, humans, platforms, regulators, and buyers act. The right question for builders and policymakers is not just “what will this model output?” It is “what system of incentives and information flows will make the desired outcome stable?”

Uncertainty quantification sits alongside economics and computation as a missing foundation in many AI systems, in Michael Jordan’s account. When an LLM is asked how confident it is, he says, the answer is not grounded in a developed account of uncertainty. To the best of his knowledge, the model is often imitating past language about confidence — how humans on the internet answered similar questions — rather than reasoning under uncertainty.

Conformal prediction is one way to wrap uncertainty around black-box models, and Jordan regards it as important. The source shows the Journal of Machine Learning Research tutorial by Glenn Shafer and Vladimir Vovk, and Jordan praises Vovk as “fantastic.” But he emphasizes assumptions and context. Classical p-values, he explains, are one-shot tail probabilities: under a model of the world, an observed outcome looks improbable, so perhaps the model is wrong. Repeatedly checking p-values and selecting the smallest leads to p-hacking.^†

E-values, by contrast, support a different style of evidence accumulation. Jordan describes an e-value as involving a non-negative random variable, or more generally a non-negative supermartingale, whose expectation under the null is bounded by one. Evidence can be gathered multiplicatively and monitored over time. With tools such as Ville’s inequality and optional stopping, one can do “anytime inference”: peek, adapt, gather new data, and still control error over the path.

The source also shows Jordan’s paper “E-Values Expand the Scope of Conformal Prediction,” whose abstract describes conformal prediction as a framework for distribution-free uncertainty quantification.^† A slide shown immediately afterward defines an e-value as a non-negative random variable $X\ge 0$ for a null hypothesis such that its expectation under every null parameter is at most one. The same slide states a theorem: “A contract is incentive-aligned if and only if all payoff functions are E-values.”

Jordan says this creates a tight connection between game-theoretic probability and the theory of incentives. In his view, uncertainty quantification is rarely just an error bar. It depends on the context in which evidence is gathered: a contract, a market, an experiment, a regulatory mechanism, or a social interaction.

His duck example makes the point. A “Bayesian duck” observes that one side of a lake has had twice as much grain as the other. Maximizing expected value as an individual would send the duck to the better side with probability one. But actual ducks distribute roughly two-thirds to one side and one-third to the other. Jordan interprets this as a population-level equilibrium. If all ducks went to the same side, they would miss resources. The relevant uncertainty is not just the individual’s uncertainty about grain; it is uncertainty in a population context.

He identifies other kinds of uncertainty as well. Information asymmetry arises when one party knows things another does not and may never fully reveal them. Provenance uncertainty concerns where data came from, how old it is, and whether it should still be trusted. If a doctor cites medical outcomes data from ten years ago, Jordan says, the confidence interval should widen. In current systems, data often flow without metadata being quantitatively incorporated into uncertainty estimates. Humans do some of this informally: they discount old information, ask someone who likely knows the local context, and adjust behavior in social settings. LLMs, he argues, do essentially none of this in a principled way.

The restaurant-and-tomatoes example gives his widest account of uncertainty reduction. If he had to forage for tomatoes every day to run a pizza restaurant, he could not plan reliably. A market reduces that uncertainty because someone else has done the foraging and supply has stabilized. Markets mitigate uncertainty not because a single experimenter designed an optimal sampling plan, but because incentives for exploration, exploitation, production, and exchange create a reliable substrate on which others can build.

The broader intellectual foundation Michael Jordan wants for the AI era is a triangle with statistics, computer science, and economics at its vertices, connected by econometrics, machine learning, and algorithmic game theory. The visual shown on screen is titled “Three Foundational Disciplines.” It places Statistics, Economics, and Computer Science at the corners of a green triangle, with econometrics, machine learning, and algorithmic game theory along the edges.

Jordan says he thinks less in terms of disciplines than in terms of thinking styles. Computational thinking, associated in the discussion with Jeannette Wing, brought ideas such as modularity, abstraction, APIs, algorithms, and limits of computation beyond computer science. The source shows Wing’s “Computational Thinking” article, whose visible subheadline describes it as “a universally applicable attitude and skill set everyone, not just computer scientists, would be eager to learn and use.” Jordan endorses that ambition.^†

But he says many important algorithms and systems do not arise from computational thinking alone. They arise from inferential thinking — how to gather data, quantify uncertainty, make predictions about things not yet observed, and control errors — and from economic thinking — how incentives, institutions, legal structures, strategic behavior, and social welfare shape outcomes.

Together, he says, these three thinking styles form a platform for training the next generation. Computational algorithms and optimization can produce LLMs. That is valuable. But by themselves they do not provide the context around the model: who owns data, who is paid, how privacy is traded off against utility, how evidence is gathered, how incentives are aligned, how regulators act, how users express bottom-up preferences, or how social welfare is evaluated.

Jordan calls this triangle “the liberal arts of the era.” He acknowledges that humanities colleagues may disagree with relocating the core in that way. His claim is narrower and tied to the problems under discussion: the central intellectual issues of the era involve data, compute, inference, incentives, institutions, and social responsibility. A serious AI education, in his view, cannot be only about optimization and model building. It must train people to design systems in which prediction, uncertainty, markets, and human goals meet.